Istanbul-based team matches frontier AI labs on knowledge-work benchmark at a fraction of the cost

ISTANBUL, TURKEY – November 14, 2026 – A small team based in Istanbul has built an AI product that matches frontier models on real professional knowledge-work tasks while operating at under a tenth of the cost of the largest models on the JobBench leaderboard.

Called Loxi, the free AI product is designed for work such as spreadsheets, legal memos, financial analysis and project planning. Unlike the frontier systems it was measured against, Loxi does not rely on a large general-purpose model. It uses a small model trained specifically for knowledge work.

On JobBench, Loxi scored 58.2%, placing it level with Claude Fable 5 and ahead of GPT-5.6 SOL, Kimi K3, Qwen 3.8 Max and Claude Opus 4.8. All 65 tasks in the main test split produced completed deliverables, with no time-outs, refusals or empty output folders. The team reports that the product operates at under a tenth of the cost of the largest models on the leaderboard.

The method behind the result

The team spent roughly eight months working on what it calls the “harness” – the layer between an AI model and the end user that determines how a task is planned, how files are read and produced, how the agent checks its own work, and which skills it loads for different types of tasks.

Rather than focusing primarily on increasing model size, the team ran thousands of controlled experiments examining the fit between the model and its harness. The experiments changed one variable at a time and measured the results on real tasks.

The team says a model and harness tuned to each other performed better than either one selected independently. In this approach, the underlying model is only one part of the system responsible for producing a completed professional deliverable.

Independent results cited in the supplied research materials point to a similar effect. LangChain changed only its agent runtime and moved from thirtieth to fifth place on Terminal Bench 2.0, increasing its score from 52.8% to 66.5% without changing the underlying model.

A separate study cited in the materials found that holding the base model fixed while redesigning the harness increased pass@1 from 69.7% to 77.0%.

Why domain-specific models are gaining ground

The Loxi result comes as companies increasingly explore AI systems designed around particular categories of work rather than relying exclusively on general-purpose frontier models.

Gartner has projected that more than half of enterprise generative AI deployments will be domain-specific by 2027, compared with about 1% in 2024.

Finance and law offer early examples of this approach. According to the supplied materials, Bridgewater Associates fine-tuned a model on how its own investment managers operate and beat frontier models at one-fourteenth of the cost. Harvey AI fine-tuned Kimi 2.6 on a legal benchmark and reported roughly a 40% improvement at one-eleventh of the cost.

Loxi follows a similar pattern. The product is free, its beta always runs the latest experimental build, and user feedback feeds directly into the team’s model and harness research.

Turkey’s AI push meets a product moment

The result comes alongside Turkey’s National AI Action Plan for 2026 to 2030, issued by presidential decree in August 2026.

According to the supplied materials, the plan targets 1 gigawatt of data-center capacity, at least $10 billion in private investment, AI literacy training for 5 million people, and 10,000 advanced specialists by 2030. It also calls for developing and exporting sector-specific language models in health, energy and smart manufacturing.

Turkey’s startup ecosystem has moved in a similar direction. The country reached eight unicorns in 2026, while the number of candidates in the Turcorn 100 program nearly doubled from 23 in March 2025 to 43 in the first half of 2026.

Ankara-based Herdr is also cited in the supplied materials as having raised $6 million in seed funding for an open-source AI agent runtime for terminals.

What it means

A July 2026 Bank of America report described the AI race as increasingly concentrated around the United States and China. Outside those two countries, the report named South Korea as the strongest candidate, followed by the UAE.

Loxi’s result points to another dimension of AI competition: the ability to combine specialized models with software engineering and task-specific infrastructure.

Benchmark tables can credit the underlying model for work that is also shaped by the surrounding system. How a task is planned, how files are handled, which capabilities are selected, and how an agent verifies its output can all affect whether a system produces a usable deliverable.

The same question is increasingly relevant to the changing knowledge-work economy. PwC’s Global AI Jobs Barometer, published in June 2026, found the labor market splitting into two tracks, with specialized roles growing twice as fast as democratized ones and showing 42% faster wage growth.

JobBench measures the type of work at the center of that shift: tasks that can be routine and repetitive while still requiring close attention and substantial reasoning.

For founders and business leaders watching AI costs rise, the result offers a different way to think about AI performance. The largest model is not necessarily the only path to capable professional AI systems; model specialization, harness design and the fit between the two can also play a significant role.

About Loxi

Loxi is a free AI product for knowledge work developed by a team based in Istanbul, Turkey. The product is built around a small model trained specifically for professional tasks and a purpose-engineered harness layer rather than a large general-purpose model.

Sources

Loxi is level with Claude Fable 5 and ahead of GPT-5.6 SOL on JobBench – August 31, 2026

Loxi is back in the top group on JobBench – September 16, 2026

Media Contact
Company Name: LOXIGO YAZILIM TEKNOLOJI ROBOTIK VE ELEKTRONIK SANAYI TICARET LIMITED SIRKETI
Contact Person: Ms. Buket Sahinbas
Email: Send Email
Country: Turkey
Website: www.loxigo.com

Media gallery